View Full Microsoft AI-103 Exam Dumps and Practice Test Dumps.

 

Question 301

Which capability allows an AI agent to retrieve current information from a business system?

1. Tool calling
2. OCR
3. Image classification
4. Text-to-speech

Correct Answer: 1

Explanation

Tool calling allows an AI agent to interact with external systems and retrieve information that is not available in its current model context. For example, an agent can call a business API to retrieve a customer’s order status or account information. The application executes the requested tool and returns its result to the agent. Developers should authenticate requests, enforce authorization, validate tool arguments, and handle failures appropriately. Tool access should also follow least-privilege principles so agents cannot perform unnecessary or unauthorized operations.

Question 302

What does OCR stand for?

1. Online Content Retrieval
2. Optical Character Recognition
3. Object Classification Response
4. Output Context Retrieval

Correct Answer: 2

Explanation

OCR stands for Optical Character Recognition. It is a technology used to extract machine-readable text from images and scanned documents. OCR is useful when documents contain text that cannot be directly selected or processed as digital text. Applications can use OCR as part of document processing, search, information extraction, and automation workflows. The quality of OCR results can depend on factors such as image resolution, orientation, handwriting, font style, and background quality. Developers should validate extracted text when it will drive important business processes.

Question 303

Why is semantic ranking useful in search applications?

1. It stores secrets securely
2. It converts audio to text
3. It can place more relevant results higher
4. It creates user identities

Correct Answer: 3

Explanation

Semantic ranking can improve the ordering of search results by considering the meaning and context of the query and retrieved content. This can help distinguish highly relevant results from documents that contain similar keywords but do not actually answer the user’s question. In RAG applications, better ranking can provide a language model with more useful context. Developers should evaluate ranking quality using representative queries and expected results. Semantic ranking does not replace access controls, filtering, or other security mechanisms that determine which documents a user is allowed to retrieve.

Question 304

Which approach can provide an AI model with company information without retraining it?

1. Image classification
2. RAG
3. Speech synthesis
4. Object detection

Correct Answer: 2

Explanation

Retrieval-augmented generation can provide an AI model with relevant company information without requiring the underlying model to be retrained whenever documents change. The application retrieves information from an approved knowledge source and supplies it as context for generation. This is useful for company policies, product documentation, support content, and other frequently updated information. Developers should maintain the knowledge source, monitor indexing freshness, enforce access controls, and evaluate retrieval quality. RAG does not automatically guarantee accurate answers if the retrieved information is incomplete or incorrect.

Question 305

What is the purpose of a model evaluation dataset?

1. Measure performance consistently
2. Store passwords
3. Increase network speed
4. Generate audio

Correct Answer: 1

Explanation

A model evaluation dataset contains representative examples that can be used to measure the performance of an AI system consistently. Developers can use the same dataset to compare different models, prompts, retrieval configurations, or application versions. Evaluation examples should reflect realistic inputs and important edge cases. Depending on the application, measurements may include accuracy, relevance, groundedness, safety, or structured-output compliance. A well-designed evaluation dataset helps identify regressions and provides evidence for technical decisions rather than relying only on informal testing.

Question 306

Which search method is useful when both exact terms and semantic meaning matter?

1. OCR
2. Hybrid search
3. Speech recognition
4. Sentiment analysis

Correct Answer: 2

Explanation

Hybrid search combines keyword-based retrieval with semantic or vector-based retrieval. Exact keyword matching can be valuable for product codes, names, identifiers, and specific terminology, while semantic retrieval can identify conceptually related information even when different words are used. Combining both approaches can improve retrieval coverage for enterprise knowledge systems. Developers should evaluate ranking behavior using representative queries because the importance of each search signal varies by workload. Hybrid search is frequently useful in RAG systems where retrieval quality directly affects the quality of generated responses.

Question 307

Which service is designed to analyze spoken audio?

1. Azure AI Speech
2. Azure AI Vision
3. Azure AI Search
4. Azure AI Document Intelligence

Correct Answer: 1

Explanation

Azure AI Speech provides capabilities for processing spoken audio, including speech-to-text, text-to-speech, and supported speech translation scenarios. Applications can use these capabilities for voice assistants, transcription, accessibility, meetings, and customer-service workflows. Developers should consider language support, audio quality, latency, speaker conditions, and expected vocabulary when designing speech solutions. Testing with realistic audio is important because background noise, accents, overlapping speakers, and microphone quality can influence results. Speech capabilities can also be combined with language or generative AI services for additional processing.

Question 308

Why should sensitive information be excluded from unnecessary prompts?

1. It reduces unnecessary data exposure
2. It increases model randomness
3. It improves microphone quality
4. It removes the need for authorization

Correct Answer: 1

Explanation

Sensitive information should only be included in a prompt when it is necessary for the requested task. Sending unnecessary personal, confidential, or business-sensitive data to a model can increase privacy and security risks. Data minimization reduces the amount of information that must be processed and potentially retained by application components. Developers should also consider access controls, logging, retention, encryption, and service-specific data-handling requirements. Prompt design should therefore provide the model with enough information to complete the task without unnecessarily exposing unrelated sensitive content.

Question 309

What is the purpose of reranking retrieved documents?

1. Convert documents into speech
2. Improve the order of candidate results
3. Remove all document metadata
4. Generate authentication tokens

Correct Answer: 2

Explanation

Reranking takes an initial set of retrieved documents and reorganizes them according to their relevance to the user’s query. A retrieval system may first produce a broad candidate set, after which a reranker identifies which results are most useful. This can improve the quality of context supplied to a generative model and reduce irrelevant information. Developers should evaluate reranking with representative queries and should consider latency and cost. Reranking does not fix fundamentally poor source data or incorrect access-control configurations.

Question 310

Which feature can help an application return responses in a predictable JSON format?

1. Structured outputs
2. Object detection
3. Speech translation
4. OCR

Correct Answer: 1

Explanation

Structured outputs allow developers to define an expected response structure for model-generated data. This is useful when an application needs predictable fields and values rather than unrestricted natural-language responses. For example, a system processing customer requests might require fields for category, priority, and summary. The application should validate the returned structure and important values before passing them to downstream systems. Structured output improves consistency but does not eliminate the need for application-level validation, error handling, or security controls.

Question 311

What does authorization determine?

1. Whether an identity is allowed to perform an action
2. Whether an image contains text
3. Whether speech can be transcribed
4. Whether a document has tables

Correct Answer: 1

Explanation

Authorization determines what an authenticated identity is permitted to access or perform. For example, an authenticated employee may be allowed to read certain documents but not delete them or access restricted records. In AI applications, authorization is especially important when agents can retrieve private information or call business tools. Developers should enforce authorization outside the model and apply least-privilege permissions. Authentication and authorization serve different purposes: authentication establishes identity, while authorization determines permitted actions or resources.

Question 312

Which capability can detect important phrases in customer feedback?

1. Object detection
2. Key phrase extraction
3. Speech synthesis
4. Vector indexing

Correct Answer: 2

Explanation

Key phrase extraction identifies important words or phrases that represent major concepts in a text. It can help applications analyze customer feedback, reviews, survey responses, reports, or support messages. Extracted phrases can support categorization, indexing, analytics, and search workflows. Key phrase extraction differs from sentiment analysis because its primary purpose is identifying significant concepts rather than determining positive or negative sentiment. Developers should test results with representative business content because important phrases and terminology can vary considerably between industries and applications.

Question 313

Why should AI-generated responses be validated in important workflows?

1. Models can produce incorrect or unsupported information
2. Validation increases storage automatically
3. Models never make mistakes
4. Validation removes the need for monitoring

Correct Answer: 1

Explanation

AI-generated responses can contain incorrect, incomplete, or unsupported information, so validation is important when outputs influence business decisions or downstream systems. Validation can include checking structured formats, required fields, business rules, source support, or other application-specific requirements. For high-impact workflows, human review may also be appropriate. Developers should not assume that a fluent response is necessarily accurate. Combining model evaluation, retrieval checks, output validation, authorization, and monitoring can provide stronger protection against incorrect or unsafe AI-generated results.

Question 314

What is a primary benefit of managed identity?

1. It can reduce the need for stored application credentials
2. It increases model temperature
3. It generates document embeddings
4. It performs OCR

Correct Answer: 1

Explanation

Managed identity can allow an Azure resource to authenticate to supported services without requiring developers to store long-lived credentials directly in application code. The identity can be assigned appropriate permissions according to the workload’s needs. This can reduce credential-management complexity and the risk of accidentally exposing secrets. Developers still need to configure authorization correctly and should follow least-privilege principles. Managed identity is especially useful for Azure-hosted applications that need to access supported resources securely without embedding service credentials in source files or configuration.

Question 315

Which service can identify objects in an image?

1. Azure AI Search
2. Azure AI Vision
3. Azure AI Speech
4. Azure AI Language

Correct Answer: 2

Explanation

Azure AI Vision provides computer vision capabilities that can be used to analyze images, including supported object-related detection scenarios. Object detection can identify objects and their locations within an image, which is useful for applications such as visual inspection, inventory analysis, and safety monitoring. Developers should evaluate detection performance with representative images because lighting, object size, image quality, occlusion, and background conditions can affect results. Vision capabilities should be selected according to the specific image-analysis requirement rather than assuming every vision task uses the same model or operation.

Question 316

What is the main purpose of rate limiting?

1. Control request frequency
2. Improve image resolution
3. Generate embeddings
4. Translate documents

Correct Answer: 1

Explanation

Rate limiting controls how frequently requests can be submitted to an application or service during a specified period. It can help protect resources from excessive traffic, unexpected request spikes, and certain forms of misuse. Rate limits can also help manage costs and prevent service quotas from being consumed too quickly. Developers should configure limits according to expected workload requirements and service capabilities. Applications should handle limit responses appropriately and avoid aggressive retry behavior that could create additional traffic. Monitoring request patterns can help identify unusual usage.

Question 317

Which capability helps an application understand the language of incoming text?

1. Language detection
2. Object detection
3. Text-to-speech
4. OCR

Correct Answer: 1

Explanation

Language detection identifies which supported language is represented by a text input. This can help applications route incoming content to appropriate translation, analysis, search, or customer-support workflows. For example, a multilingual application can detect the language before selecting a suitable processing path. Developers should verify that all required languages are supported and test realistic inputs. Very short text, mixed-language content, unusual spelling, or ambiguous words may make detection more difficult. Language detection works on text and should not be confused with speech recognition.

Question 318

What is the purpose of human approval in an AI workflow?

1. Review selected actions before they occur
2. Increase embedding dimensions
3. Replace search indexing
4. Disable authorization

Correct Answer: 1

Explanation

Human approval allows a person to review an AI-generated action before it is executed. This can be useful for sensitive or high-impact operations such as financial transactions, account changes, or actions affecting important business records. The reviewer should receive enough information to understand the proposed operation and make an informed decision. Human approval should complement, not replace, authentication, authorization, validation, and monitoring. Developers should define which operations require approval and design the workflow so that unapproved actions cannot be executed automatically.

Question 319

Which approach is useful when an AI application needs information that changes frequently?

1. Static prompts only
2. RAG with an updated knowledge source
3. Permanent hard-coded responses
4. Removing document indexing

Correct Answer: 2

Explanation

RAG with an updated knowledge source can provide a generative AI application with current information without requiring the underlying model to be retrained whenever the information changes. Documents can be updated, indexed, and retrieved when users submit questions. This is useful for policies, product catalogs, support documentation, and other changing content. Developers should maintain source freshness, remove or version outdated information, and apply access controls. Retrieval quality should also be evaluated regularly because stale or irrelevant content can lead to poor generated responses.

Question 320

What should developers monitor after changing an AI model?

1. Quality, latency, errors, and usage
2. Only the application’s logo
3. Only document colors
4. Only keyboard input

Correct Answer: 1

Explanation

After changing an AI model, developers should monitor important operational and quality metrics to identify unexpected effects. These can include response quality, groundedness, latency, token usage, error rates, throughput, and safety-related outcomes where applicable. A new model may behave differently even when the application code remains unchanged. Evaluation datasets can provide controlled comparisons before and after the change, while production monitoring can reveal real-world behavior. Combining pre-deployment evaluation with post-deployment monitoring helps identify regressions and operational issues early.